It sounds like the machine that the DB VM was running on was being taken offline or restarted for maintenance. The default behaviour for GCE is to live-migrate the VM to another machine. So I guess Cloud SQL uses the default here. (It may well be the best option as failover isn't instant either.) Live migration is usually much faster than 90s at well but if you are making heavy RAM updates that could definitely slow it down.

Either way I agree that some full-stack integration is needed on GCPs part to at least get that into the maitnaince log. It would also be nice to make most of these happen during the maitnaince window but IIUC they don't always have 24h notice of a machine reboot.